Thanks for the proposal! I like this idea as it gives Flink's adaptive batching processing more room to imagine and optimize.
So, +1 from my side. I just have two questions: 1. `StreamGraphOptimizationStrategy` is a reasonable abstract, I'd like to know what built-in strategy implementations you have in mind so far? 2. For the so-called pending operators, can we show it in different colors on the UI. Best regards, Weijie Zhu Zhu <reed...@gmail.com> 于2024年7月17日周三 16:49写道： > Thanks Junrui for the updates. The proposal looks good to me. > With the stream graph added to the REST API result, I think we are > also quite close to enable Flink to expand a job vertex to show its > operator-chain topology. > > Thanks, > Zhu > > Junrui Lee <jrlee....@gmail.com> 于2024年7月15日周一 14:58写道： > > > Hi Zhu, > > > > Thanks for your feedback. > > > > Following your suggestion, I have updated the public interface section of > > the FLIP with the following additions: > > > > 1. UI: > > The job topology will display a hybrid of the current JobGraph along with > > downstream components yet to be converted to a StreamGraph. On the > topology > > graph display page, there will be a "Show Pending Operators" button in > the > > upper right corner for users to switch back to a job topology that only > > includes JobVertices. > > > > 2. Rest API: > > Add a new field "stream-graph-plan" will be added to the job details REST > > API, which represents the runtime Stream graph. The field "job-vertex-id" > > is valid only when the StreamNode has been converted to a JobVertex, and > it > > will hold the ID of the corresponding JobVertex for that StreamNode. > > > > For further information, please feel free to review the public interface > > section of FLIP-469 > > > > Best, > > Junrui > > > > Zhu Zhu <reed...@gmail.com> 于2024年7月15日周一 10:29写道： > > > > > +1 for the FLIP > > > > > > It is useful to adaptively optimize logical execution plans(stream > > > operators and > > > stream edges) for batch jobs. > > > > > > One question: > > > The FLIP already proposed to update the REST API & Web UI to show > > operators > > > that are not yet converted to job vertices. However, I think it would > be > > > better if Flink can display these operators as part of the graph, > > allowing > > > users to have an overview of the processing logic graph at early stages > > of > > > the job execution. > > > This change would also involve the public interface, so instead of > > > postponing > > > it to a later FLIP, I prefer to have a design for it in this FLIP. > WDYT? > > > > > > Thanks, > > > Zhu > > > > > > Junrui Lee <jrlee....@gmail.com> 于2024年7月11日周四 11:27写道： > > > > > > > Hi devs, > > > > > > > > Xia Sun, Lei Yang, and I would like to initiate a discussion about > > > > FLIP-469: Supports Adaptive Optimization of StreamGraph. > > > > > > > > This FLIP is the second in the series on adaptive optimization of > > > > StreamGraph and follows up on FLIP-468 [1]. As we proposed in > FLIP-468 > > to > > > > enable the scheduler to recognize and access the StreamGraph, in this > > > FLIP, > > > > we will propose a mechanism for adaptive optimization of StreamGraph. > > It > > > > allows the scheduler to dynamically adjust the logical execution plan > > at > > > > runtime. This mechanism is the base of adaptive optimization > > strategies, > > > > such as adaptive broadcast join and skewed join optimization. > > > > > > > > Unlike the traditional execution of jobs based on a static > StreamGraph, > > > > this mechanism will progressively determine StreamGraph during > runtime. > > > The > > > > determined StreamGraph will be transformed into a specific JobGraph, > > > while > > > > the indeterminate part will allow Flink to flexibly adjust according > to > > > > real-time job status and actual input conditions. > > > > > > > > Note that this FLIP focuses on the introduction of the mechanism and > > does > > > > not introduce any actual optimization strategies; these will be > > discussed > > > > in subsequent FLIPs. > > > > > > > > For more details, please refer to FLIP-469 [2]. We look forward to > your > > > > feedback. > > > > > > > > Best, > > > > > > > > Xia Sun, Lei Yang and Junrui Lee > > > > > > > > [1] > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-468%3A+Introducing+StreamGraph-Based+Job+Submission > > > > [2] > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/FLINK/FLIP-469%3A+Supports+Adaptive+Optimization+of+StreamGraph > > > > > > > > > >
